Bayesian pollution source identification via an inverse physics model

Abstract

The behavior of air pollution is governed by complex dynamics in which the air quality of a site is affected by the pollutants transported from neighboring locations via physical processes. To estimate the sources of observed pollution, it is crucial to take the atmospheric conditions into account. Traditional approaches to building empirical models use observations, but do not extensively incorporate physical knowledge. Failure to exploit such knowledge can be critically limiting, particularly in situations where near-real-time estimation of a pollution source is necessary. A Bayesian method is proposed to estimate the locations and relative contributions of pollution sources by incorporating both the physical knowledge of fluid dynamics and observed data. The proposed method uses a flexible approach to statistically utilize large-scale data from a numerical weather prediction model while integrating the dynamics of the physical processes into the model. This method is illustrated with a real wind data set.